Knowledge discovery for time series

My thesis investigates the use of machine learning methods for analysis of economic and financial time series. Since structural models in economics and finance are known to have limited predictive power, I study a data driven, time series approach to knowledge discovery in these domains. The ultimate goal of building predictive models of such time series is to support decision making in areas such as business, investing, and government policy. Machine learning offers powerful tools for forecasting and decision making. Supervised learning methods can be used to develop forecasting models of economic series that can aid in decision support. Reinforcement learning methods can produce systems capable of making investment decisions. Hence, my thesis consists of two main investigations: a study of methods for predicting macroeconomic and financial time series, and a study of extensions to a reinforcement learning algorithm for constructing financial decision systems. In the forecasting project, I develop a supervised training methodology for models that predict challenging macroeconomic and financial time series. I compare the performance of linear and nonlinear networks with a diverse set of standard linear benchmark models. While some advantage is obtained from the use of nonlinear networks for certain of these time series, a key result is that linear network models trained with stochastic, nonlinear neural network learning algorithms can achieve greatly improved performance over the benchmark methods on most of the data sets. The second topic investigated is enhancements to the Recurrent Reinforcement Learning (RRL) algorithm. The RRL approach to trading system design has been shown to be effective at learning strategies that directly maximize financial objective functions, and also has been shown to outperform approaches based on supervised learning on artificial data sets. In my work, I investigate several significant extensions of RRL: to incorporate downside risk measures, to compare the RRL policy approach to an alternate RL value function approach, to extend the approach to portfolio management, and to conduct simulation studies on a number of artificial and real data sets, including an S&P-500 asset allocation system and a high frequency foreign exchange trader.